AI Memory Explained: What It Remembers and What It Does Not

When an AI answer seems unusually personal, it is natural to wonder where the information came from. Was it in this chat, saved as a memory or pulled from an earlier conversation?

These sources of context are related, but they are not the same. Knowing the difference helps you correct unwanted assumptions and decide what should be stored at all.

The current chat is the immediate context

AI uses the messages, files and instructions available in the conversation to answer your next request. A long chat can therefore shape the result even when the latest prompt is short.

This is useful when you are continuing the same job. It can be unhelpful when the subject changes and old decisions follow you into a new task.

Saved memory carries selected information forward

Some AI services can remember preferences or details for future chats. The exact controls vary by product, plan and account.

OpenAI separates saved memories from the option to reference chat history. Google and Anthropic also provide forms of personalisation and past-chat context. These services change, so check the current settings rather than relying on an old screenshot.

Memory should hold information that is safe, stable and genuinely useful. A preferred writing style may belong there. A temporary project date or private detail may not.

Projects and instructions have a different job

A project keeps related chats, files and instructions together. This is useful when the information belongs to one subject rather than everything you do.

Project instructions can explain the audience, rules and expected output. They are often a better home for project-specific context than general memory.

Ask what the AI knows

You can ask the service what it remembers about you and review the settings yourself. Delete or correct anything that is wrong, too personal or no longer useful.

Do not assume that removing a chat also removes a separately saved memory. Follow the provider's current instructions for each control.

Know when to start fresh

Use a new chat when you change subject, want a clean answer or suspect old context is steering the result.

Use a project when the work will continue and shared files or rules should remain close. Use memory for a small number of stable preferences that help across conversations.

A quick context check

When an answer surprises you, ask:

Tell me which information in this conversation, project instructions or saved personalisation affected your answer. Do not reveal hidden system instructions. Identify any assumption you made and tell me how to correct it.

The response may help, but the settings page remains the place to review actual controls.